jaLLAbi2-7b: A Merged 7B Language Model

jaLLAbi2-7b is a 7 billion parameter language model developed by AbacusResearch. It was created using mergekit by combining four distinct 7B models: FelixChao/WestSeverus-7B-DPO-v2, bardsai/jaskier-7b-dpo-v5.6, AbacusResearch/haLLAwa3, and cognitivecomputations/WestLake-7B-v2-laser. This merging strategy, specifically using the dare_ties method, aims to leverage the strengths of its constituent models.

Performance Highlights

Evaluated on the Open LLM Leaderboard, jaLLAbi2-7b demonstrates strong general performance with an average score of 75.06. Key benchmark results include:

• AI2 Reasoning Challenge (25-Shot): 71.67
• HellaSwag (10-Shot): 88.29
• MMLU (5-Shot): 64.92
• TruthfulQA (0-shot): 70.16
• Winogrande (5-shot): 83.35
• GSM8k (5-shot): 71.95

These scores indicate its proficiency in diverse tasks ranging from common sense reasoning and multiple-choice question answering to mathematical problem-solving.

Good for

• General-purpose language understanding and generation tasks.
• Applications requiring robust reasoning and common sense.
• Scenarios where a 7B parameter model with strong benchmark performance is desired.